1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/context_tracking.h>
4 #include <linux/entry-common.h>
5 #include <linux/highmem.h>
6 #include <linux/livepatch.h>
7 #include <linux/audit.h>
8 #include <linux/tick.h>
12 #define CREATE_TRACE_POINTS
13 #include <trace/events/syscalls.h>
15 /* See comment for enter_from_user_mode() in entry-common.h */
16 static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
18 arch_check_user_regs(regs);
19 lockdep_hardirqs_off(CALLER_ADDR0);
21 CT_WARN_ON(ct_state() != CONTEXT_USER);
24 instrumentation_begin();
25 trace_hardirqs_off_finish();
26 instrumentation_end();
29 void noinstr enter_from_user_mode(struct pt_regs *regs)
31 __enter_from_user_mode(regs);
34 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
36 if (unlikely(audit_context())) {
37 unsigned long args[6];
39 syscall_get_arguments(current, regs, args);
40 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
44 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
50 * Handle Syscall User Dispatch. This must comes first, since
51 * the ABI here can be something that doesn't make sense for
52 * other syscall_work features.
54 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
55 if (syscall_user_dispatch(regs))
60 if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
61 ret = arch_syscall_enter_tracehook(regs);
62 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
66 /* Do seccomp after ptrace, to catch any tracer changes. */
67 if (work & SYSCALL_WORK_SECCOMP) {
68 ret = __secure_computing(NULL);
73 /* Either of the above might have changed the syscall number */
74 syscall = syscall_get_nr(current, regs);
76 if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
77 trace_sys_enter(regs, syscall);
79 syscall_enter_audit(regs, syscall);
81 return ret ? : syscall;
84 static __always_inline long
85 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
87 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
89 if (work & SYSCALL_WORK_ENTER)
90 syscall = syscall_trace_enter(regs, syscall, work);
95 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
97 return __syscall_enter_from_user_work(regs, syscall);
100 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
104 __enter_from_user_mode(regs);
106 instrumentation_begin();
108 ret = __syscall_enter_from_user_work(regs, syscall);
109 instrumentation_end();
114 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
116 __enter_from_user_mode(regs);
117 instrumentation_begin();
119 instrumentation_end();
122 /* See comment for exit_to_user_mode() in entry-common.h */
123 static __always_inline void __exit_to_user_mode(void)
125 instrumentation_begin();
126 trace_hardirqs_on_prepare();
127 lockdep_hardirqs_on_prepare();
128 instrumentation_end();
131 arch_exit_to_user_mode();
132 lockdep_hardirqs_on(CALLER_ADDR0);
135 void noinstr exit_to_user_mode(void)
137 __exit_to_user_mode();
140 /* Workaround to allow gradual conversion of architecture code */
141 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
143 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
145 if (ti_work & _TIF_NOTIFY_SIGNAL)
146 tracehook_notify_signal();
148 arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
151 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
152 unsigned long ti_work)
155 * Before returning to user space ensure that all pending work
156 * items have been completed.
158 while (ti_work & EXIT_TO_USER_MODE_WORK) {
160 local_irq_enable_exit_to_user(ti_work);
162 if (ti_work & _TIF_NEED_RESCHED_MASK)
165 #ifdef ARCH_RT_DELAYS_SIGNAL_SEND
166 if (unlikely(current->forced_info.si_signo)) {
167 struct task_struct *t = current;
168 force_sig_info(&t->forced_info);
169 t->forced_info.si_signo = 0;
173 if (ti_work & _TIF_UPROBE)
174 uprobe_notify_resume(regs);
176 if (ti_work & _TIF_PATCH_PENDING)
177 klp_update_patch_state(current);
179 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
180 handle_signal_work(regs, ti_work);
182 if (ti_work & _TIF_NOTIFY_RESUME)
183 tracehook_notify_resume(regs);
185 /* Architecture specific TIF work */
186 arch_exit_to_user_mode_work(regs, ti_work);
189 * Disable interrupts and reevaluate the work flags as they
190 * might have changed while interrupts and preemption was
193 local_irq_disable_exit_to_user();
195 /* Check if any of the above work has queued a deferred wakeup */
196 tick_nohz_user_enter_prepare();
198 ti_work = READ_ONCE(current_thread_info()->flags);
201 /* Return the latest work state for arch_exit_to_user_mode() */
205 static void exit_to_user_mode_prepare(struct pt_regs *regs)
207 unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
209 lockdep_assert_irqs_disabled();
211 /* Flush pending rcuog wakeup before the last need_resched() check */
212 tick_nohz_user_enter_prepare();
214 if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
215 ti_work = exit_to_user_mode_loop(regs, ti_work);
217 arch_exit_to_user_mode_prepare(regs, ti_work);
219 /* Ensure that the address limit is intact and no locks are held */
220 addr_limit_user_check();
222 lockdep_assert_irqs_disabled();
227 * If SYSCALL_EMU is set, then the only reason to report is when
228 * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP). This syscall
229 * instruction has been already reported in syscall_enter_from_user_mode().
231 static inline bool report_single_step(unsigned long work)
233 if (work & SYSCALL_WORK_SYSCALL_EMU)
236 return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
239 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
244 * If the syscall was rolled back due to syscall user dispatching,
245 * then the tracers below are not invoked for the same reason as
246 * the entry side was not invoked in syscall_trace_enter(): The ABI
247 * of these syscalls is unknown.
249 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
250 if (unlikely(current->syscall_dispatch.on_dispatch)) {
251 current->syscall_dispatch.on_dispatch = false;
256 audit_syscall_exit(regs);
258 if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
259 trace_sys_exit(regs, syscall_get_return_value(current, regs));
261 step = report_single_step(work);
262 if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
263 arch_syscall_exit_tracehook(regs, step);
267 * Syscall specific exit to user mode preparation. Runs with interrupts
270 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
272 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
273 unsigned long nr = syscall_get_nr(current, regs);
275 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
277 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
278 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
285 * Do one-time syscall specific work. If these work items are
286 * enabled, we want to run them exactly once per syscall exit with
287 * interrupts enabled.
289 if (unlikely(work & SYSCALL_WORK_EXIT))
290 syscall_exit_work(regs, work);
293 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
295 syscall_exit_to_user_mode_prepare(regs);
296 local_irq_disable_exit_to_user();
297 exit_to_user_mode_prepare(regs);
300 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
302 __syscall_exit_to_user_mode_work(regs);
305 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
307 instrumentation_begin();
308 __syscall_exit_to_user_mode_work(regs);
309 instrumentation_end();
310 __exit_to_user_mode();
313 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
315 __enter_from_user_mode(regs);
318 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
320 instrumentation_begin();
321 exit_to_user_mode_prepare(regs);
322 instrumentation_end();
323 __exit_to_user_mode();
326 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
328 irqentry_state_t ret = {
332 if (user_mode(regs)) {
333 irqentry_enter_from_user_mode(regs);
338 * If this entry hit the idle task invoke rcu_irq_enter() whether
339 * RCU is watching or not.
341 * Interrupts can nest when the first interrupt invokes softirq
342 * processing on return which enables interrupts.
344 * Scheduler ticks in the idle task can mark quiescent state and
345 * terminate a grace period, if and only if the timer interrupt is
346 * not nested into another interrupt.
348 * Checking for rcu_is_watching() here would prevent the nesting
349 * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
350 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
351 * assume that it is the first interrupt and eventually claim
352 * quiescent state and end grace periods prematurely.
354 * Unconditionally invoke rcu_irq_enter() so RCU state stays
357 * TINY_RCU does not support EQS, so let the compiler eliminate
358 * this part when enabled.
360 if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
362 * If RCU is not watching then the same careful
363 * sequence vs. lockdep and tracing is required
364 * as in irqentry_enter_from_user_mode().
366 lockdep_hardirqs_off(CALLER_ADDR0);
368 instrumentation_begin();
369 trace_hardirqs_off_finish();
370 instrumentation_end();
377 * If RCU is watching then RCU only wants to check whether it needs
378 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
379 * already contains a warning when RCU is not watching, so no point
380 * in having another one here.
382 lockdep_hardirqs_off(CALLER_ADDR0);
383 instrumentation_begin();
384 rcu_irq_enter_check_tick();
385 trace_hardirqs_off_finish();
386 instrumentation_end();
391 void irqentry_exit_cond_resched(void)
393 if (!preempt_count()) {
394 /* Sanity check RCU and thread stack */
395 rcu_irq_exit_check_preempt();
396 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
397 WARN_ON_ONCE(!on_thread_stack());
398 if (should_resched(0))
399 preempt_schedule_irq();
402 #ifdef CONFIG_PREEMPT_DYNAMIC
403 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
406 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
408 lockdep_assert_irqs_disabled();
410 /* Check whether this returns to user mode */
411 if (user_mode(regs)) {
412 irqentry_exit_to_user_mode(regs);
413 } else if (!regs_irqs_disabled(regs)) {
415 * If RCU was not watching on entry this needs to be done
416 * carefully and needs the same ordering of lockdep/tracing
417 * and RCU as the return to user mode path.
419 if (state.exit_rcu) {
420 instrumentation_begin();
421 /* Tell the tracer that IRET will enable interrupts */
422 trace_hardirqs_on_prepare();
423 lockdep_hardirqs_on_prepare();
424 instrumentation_end();
426 lockdep_hardirqs_on(CALLER_ADDR0);
430 instrumentation_begin();
431 if (IS_ENABLED(CONFIG_PREEMPTION)) {
432 #ifdef CONFIG_PREEMPT_DYNAMIC
433 static_call(irqentry_exit_cond_resched)();
435 irqentry_exit_cond_resched();
438 /* Covers both tracing and lockdep */
440 instrumentation_end();
443 * IRQ flags state is correct already. Just tell RCU if it
444 * was not watching on entry.
451 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
453 irqentry_state_t irq_state;
455 irq_state.lockdep = lockdep_hardirqs_enabled();
458 lockdep_hardirqs_off(CALLER_ADDR0);
459 lockdep_hardirq_enter();
462 instrumentation_begin();
463 trace_hardirqs_off_finish();
465 instrumentation_end();
470 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
472 instrumentation_begin();
474 if (irq_state.lockdep) {
475 trace_hardirqs_on_prepare();
476 lockdep_hardirqs_on_prepare();
478 instrumentation_end();
481 lockdep_hardirq_exit();
482 if (irq_state.lockdep)
483 lockdep_hardirqs_on(CALLER_ADDR0);